All-Optical Neuron Breaks the Nanosecond Barrier Using Tellurium Phase Transition

Key Takeaways

• Tellurium melts in 260 ps, enabling sub‑nanosecond activation.
• Threshold energy as low as 0.4 pJ per neuron.
• Footprint shrinks to under 5 µm², allowing dense integration.
• Demonstrated 96.4% MNIST accuracy, 100× faster than optoelectronic.
• Over 5 billion switching cycles, surpassing GST durability.

Summary

Researchers have demonstrated an all‑optical neuron built from a thin tellurium film that melts in under 260 picoseconds, breaking the nanosecond barrier for photonic activation. The device operates with threshold energies as low as 0.4 picojoules and occupies less than 5 µm², enabling dense on‑chip integration. Integrated into a three‑layer photonic network, it classified MNIST digits with 96.4% accuracy, processing 50 images in roughly 20 µs—about 100 times faster than comparable optoelectronic systems. Durability tests showed over five billion switching cycles, far exceeding prior phase‑change materials.

Pulse Analysis

Photonic neural networks promise AI inference at the speed of light, but their progress has been throttled by the nonlinear activation stage, which traditionally required an electronic detour. Thermo‑optic and free‑carrier dispersion schemes have delivered micro‑ to nanosecond response times, while phase‑change materials like GST introduced reset overheads and limited durability. The core challenge has been finding a material that can modulate light instantly and revert without external control, a prerequisite for keeping the entire inference pipeline in the optical domain.

Tellurium’s unique solid‑to‑liquid transition under femtosecond laser pulses provides that missing piece. When a pulse deposits roughly 0.4 pJ, the film locally melts, sharply increasing its extinction coefficient and producing a sub‑nanosecond drop in transmission. The molten region then self‑quenches, recrystallizing within 260 ps and readying the neuron for the next spike. This mechanism yields an ultra‑compact footprint (1.5‑4 µm²) and energy efficiency that outpaces GST‑based designs by two orders of magnitude, while delivering 100× faster activation than the best prior all‑optical neurons.

The practical impact extends beyond a laboratory curiosity. By integrating tellurium neurons with GST‑based photonic synapses, researchers achieved a compute density of 3.6 TOPS mm⁻² and an energy efficiency of 13.6 TOPS W⁻¹—metrics that rival or exceed contemporary electronic AI accelerators. As pulse widths shrink toward sub‑picosecond regimes and device geometries are refined, response times could dip below 35 ps, effectively erasing the latency gap between neurons and synapses. This positions all‑optical processors as viable contenders for next‑generation data‑center and edge AI workloads, where bandwidth, latency, and power constraints dominate design decisions.